Electrocardiogram (ECG) signals are used to detect the health status of the heart, providing an important basis for the prevention and diagnosis of cardiovascular diseases. However, ECG signals are susceptible to environmental and equipment-related influences, which can obscure the characteristic information within the signals. Removing noise from ECG signals is an urgent problem. This paper proposes a noise-reduction method for low-frequency ECG signals using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Tuna Swarm Optimization (TSO), and Stacked Sparse Autoencoder (SSAE), named CEEMDAN-TSO-SSAE. The TSO algorithm optimizes three parameters of the CEEMDAN algorithm: Noise Standard Deviation, Number of Realizations, and Maximum Iterations. These optimized parameters are then applied to decompose the ECG signals using CEEMDAN, and the Intrinsic Mode Functions (IMFs) are obtained. The correlation coefficient method is used to screen the IMFs, excluding modal components that do not meet the threshold. Finally, each effective IMF is denoised separately using the SSAE algorithm, and the denoised effective IMFs are used for signal reconstruction. To validate the effectiveness of CEEMDAN-TSO-SSAE, its performance is compared with wavelet packet decomposition, Empirical Mode Decomposition, TSO-based Variational Mode Decomposition, and a Denoising Autoencoder algorithm. The noise-reduction method using CEEMDAN-TSO-SSAE achieves the highest Signal-to-Noise Ratio (SNR) of 19.88 and the lowest Mean Squared Error (MSE) of 0.02. In tests using real signals with baseline drift, the CEEMDAN-TSO-SSAE method again produces the highest SNR (20.25) and the lowest MSE (0.01). The results demonstrate that the proposed method outperforms the comparative algorithms, effectively eliminating complex noise in ECG signals while preserving the useful components.